Teach Robots to Learn from the Real World

Infinite Hands is looking for an ML Researcher - Robot Learning to build, adapt, and critically evaluate deep-learning models for robotics. The target is vision-language-action learning: language instructions, camera observations, and robot state inform action sequences for two-handed manipulation.

This is hands-on model work. You will investigate why a model succeeds or fails, change the model or its training inputs, and test whether the change holds up beyond an offline metric. Experienced deep-learning practitioners from other domains are welcome if they can show substantive model work and sound experimental judgment. Prior robot-learning experience is valuable, but it is not a prerequisite.

Infinite Hands is building embodied AI for aerospace manufacturing. Its aerospace design-partner relationship gives the team access to an active beta factory environment where research can be tested against real workflows. A technical founder brings experience from approximately 12 years at Apple across Siri, applied AI, robotics, and future-device systems.

The opportunity is to connect research questions to physical tasks and deployment evidence. Factory data and robot behavior will expose problems that a clean benchmark may miss; those observations should shape the next experiment.

• Train and evaluate models using language, camera observations, robot state, and action sequences for bimanual manipulation.
• Examine manipulation datasets for quality issues, uneven coverage, misleading distributions, and shifts between training and deployment.
• Change model architecture, inputs, outputs, training choices, or internal diagnostics when the evidence suggests a better way to represent or solve the task.
• Design comparisons and failure analyses that separate meaningful behavior improvements from attractive but misleading metrics.
• Move promising work from offline experiments into simulation and, where appropriate, real-robot validation. Use what each stage reveals to refine the model and evaluation.
• Explain findings clearly enough that the team can decide what to try, collect, or test next.

A useful result could be a model improvement, a convincing failure diagnosis, or evidence that a proposed change does not transfer. We care about the reasoning behind the experiment: what behavior you were trying to change, which data or model choice you made, what you compared, and what the result supports.

Research depth and implementation both matter. You should be comfortable working directly with models and training or evaluation code, inspecting data and outputs, and revising a hypothesis when robot behavior disagrees with an offline score. Success looks like useful behavior, defensible experiments, and findings that guide the next model or data decision. The team will discuss near-term priorities with you.

• Attributable hands-on work developing, training, adapting, or evaluating deep-learning models. Tell us what you personally changed and why.
• Practical experimental judgment: you can choose a meaningful comparison, inspect failures, and distinguish an actual model improvement from noise or a metric artifact.
• Enough model and data intuition to transfer your experience toward vision-language-action learning for robots, including questions you would need to test rather than assume away.
• Ability to build and test working models, explain limitations honestly, and collaborate on what the evidence means for a physical task.
• Fluent spoken and written English.

Industry model development, research engineering, applied science, and academic research can all provide this evidence. We will look at the work itself rather than require a particular title, school, degree, or number of years.

• Robot learning, manipulation, vision-language-action models, multimodal models, imitation learning, or learned robot policies.
• Diagnosing dataset quality or distribution shift and building evaluations that reveal model behavior.
• Simulation or physical-robot experiments that connect a learned model to an action task.
• A relevant PhD or research track record backed by direct model and experiment contributions.
• PyTorch, OpenPI, OpenVLA, LeRobot, or manipulation hardware. Equivalent tools and methods are welcome.

Deep-learning experience outside robotics is an accepted route. If you have improved or diagnosed learned models in another domain, show how you made the model and experimental choices; we can discuss how that experience transfers to robot data and action.

The role is based in Miami, with on-site or remote work possible and travel for key events. The team will agree on the practical location and travel arrangement during the process.

Candidates need existing authorization to work in the United States without sponsorship. Some customer or defense-related facilities may have additional access requirements, which will be resolved during the process.

The annual base salary budget is $160,000-$200,000, plus equity. The individual equity grant and package will be discussed during the process.

You will work on a research problem with a path to physical evaluation in an active aerospace manufacturing environment. The value of the work is visible in what a robot can do, what fails, and what the team learns from each experiment.

• You want to work only on model-serving infrastructure, API integrations, or AI applications without changing and evaluating learned models themselves.
• You prefer benchmark scores alone to inspecting data, failures, and behavior under new conditions.
• You want a fully specified research plan before testing ideas against evidence from simulation or real systems.
• Potential travel for key events does not work for you.

The process will include discussion of your actual model work: what you changed, which experiments you chose, how you evaluated results, and what you learned from failures. The team will explain the format and expectations for any technical conversation or assessment before it is scheduled.